Motion graphics generation benchmark from Figma and Stanford
Expert benchmark with programmatic verification to evaluate how LLMs animate vector graphics.
Hugging Face datasets with plain-language descriptions: what's inside and what models to train on them. 4,788 in the catalog so far, with new ones added every hour.
Answered by Gemini 4 · the chat is kept only in this tab
Expert benchmark with programmatic verification to evaluate how LLMs animate vector graphics.
Small test set of 218 eye images with manual segmentation masks.
Dish images with metadata for food recognition, classification, and nutritional analysis.
A collection of image and text datasets for evaluating models on multimodal graphs.
Recordings from two stereo endoscopes (~48 GB) for depth estimation, pose, and 3D reconstruction in surgery.
The first large-scale dataset of ~2M vector Lottie animations for training multimodal graphics generation.
Pairs of mathematical formula images with LaTeX code for training and evaluating OCR.
123k most difficult image-text examples from MMFineReason where Qwen3-VL-4B-Thinking consistently fails.
Nearly 130 thousand chemical reactions and templates extracted by an agent from scientific illustrations.
A 406-step Claude Code session log generating Blender code for a pelican on a bicycle animation.
A line dataset of handwritten Chinese text collected from over a thousand authors.
Collection of 1,030 Google 3D object scans with textures and meshes.
A set of 248 corporate document images for training classifiers.
Multiple-choice visual question dataset for testing 3D spatial understanding in multimodal models.
A set of 110 complex dialogues and over 1600 questions across image series for evaluating multimodal models.
Millions of image and text examples across six languages for training multimodal models.
7.3 million multimodal examples from 1 million websites for training web agents and VLM models.
A dataset of 308K synthetic pixel art images paired with text descriptions.
An exam-based benchmark for evaluating LLMs in the Japanese financial domain.
10 to 100 million image-text pairs for precise word-to-object mapping in images.
Geospatial imagery and mask dataset based on 31 landslide inventories from 21 seismic regions for disaster monitoring.
Dataset of 3,704 chess openings with ECO classification for game analysis.
A VQA dataset based on Localized Narratives featuring image question-answer pairs.
Test set of image-question pairs in 23 languages for evaluating multilingual vision-language models.